Brian Dillon Receives NSF Grant to Explore AI and Human Language Processing
Content
Brian Dillon, professor in the Department of Linguistics at UMass Amherst, has been awarded a research grant from the National Science Foundation (NSF) to study how artificial intelligence (AI) systems and humans differ in the way they process language. The four-year project, titled "Collaborative Research: Semantic Focusing: Controlling LM Interpretations for Human-Model Alignment," will begin in fall 2025. Dillon will lead the work at UMass in collaboration with Tal Linzen, associate professor of linguistics and data science at New York University.
Together, UMass Amherst and NYU have received approximately $1.2 million in funding for the project, with $432,656 awarded to UMass. The research aims to explore how large language models interpret meaning, with the goal of bringing their processing more in line with how humans understand language.
Dillon, a psycholinguist who directs the Computational Sentence Processing Lab at UMass Amherst, is particularly interested in how grammar and working memory interact during language comprehension. His research combines psycholinguistic and computational approaches to investigate sentence processing in adults.
We spoke with Dillon about his research, recent projects, and his experience in the Department of Linguistics at UMass Amherst.
Tell us about this research.
We are hoping to better understand some of the most important ways in which AI systems and humans differ in how they process language. The way we want to do this is to better understand the way that AI systems encode aspects of the meaning conveyed by simple language. Armed with a better understanding of this, we hope to also explore tools for manipulating and editing aspects of the meaning these models encode. Ultimately, with this research, we’re hoping to bring the models’ language processing more in line with how we think humans process language—that is, attempting to identify the best interpretation of linguistic input as quickly as possible. We also aim to understand if the models we develop using this approach help us capture a range of puzzling and important findings about human language processing that remain unexplained.
Who are you working with?
This project is a collaboration with Tal Linzen of New York University. Tal Linzen is an associate professor of linguistics and data science, and a research scientist at Google.
How might this project shape or influence your future research or the field more broadly?
Contemporary AI tools are ultimately based in psycholinguistic theories of how humans process and learn from language, and these models have now begun to profoundly shape psycholinguistic research in return. We’re hoping that this research continues this positive feedback cycle between basic science on language processing and language technologies. Further insights about how humans process language could yield new insights into how to structure artificial neural language models, and those models could in turn help researchers focused on understanding humans make sense of puzzling data.
Tell us about your current work and the questions that have been especially exciting to you lately.
One big idea in cognitive science and artificial intelligence is that prediction is an important component of intelligent systems. To take one example, many foundation language models in AI use prediction as a key learning signal: they try to predict the next word in a text given the words that precede it, and they use failed prediction as a learning signal. That is, the models try to learn the best representations that let them do better at predicting upcoming words. This gets you surprisingly far! It’s also quite well established that we humans do something similar when we read: we try to predict the next word, and we’re very sensitive to the probability of a word in text, suggesting we’re also making predictions about what’s likely to come next.
A natural question is how far you can take this simple idea, and one thing we’ve shown in recent work from the group is that prediction alone doesn’t seem to be able to account for all the difficulty readers face when encountering unexpected words. Instead, it seems that something about the linguistic structure of a text is important in setting human readers’ expectations above and beyond simple word predictability. This is something that psycholinguists have argued for some time now, but recent advances in large-scale data collection and statistical modeling have allowed us and others to rigorously show this is the case. But there are many unresolved questions here! How and why does the grammatical structure of our language influence language processing? Is this an important component of learning, like next-word prediction is? And are there any insights we can gain from studying human language processing behavior that will let us develop more robust, or more data-efficient, AI systems? These are some of the overarching questions that excite us.
What stands out about UMass and your department?
Our Linguistics department here has long been recognized as a leading research group in linguistics. Our department is a tight knit, research-oriented community that so many researchers from all over the world come to spend time in. I think being part of that community, and sharing in the bridge that it makes to the wider global linguistics research community, is the standout feature of the Linguistics department here.
What advice would you give students?
It’s a very exciting time to be in language science and linguistics—things are moving at a faster pace than I would’ve imagined even just five or six years ago. My advice for students in this moment is to keep an open mind, and be ready to find inspiration or insight in places that you might not have expected to find it. We’re poised to learn great things—and I think it is the students who will teach us these things—but I think we may find that what we learn looks very unlike what we might’ve expected, from the linguistics side or from the AI side.